SimpleCamera simpleCamera(const Matrix& P) {
// P = [A|a] = s K cRw [I|-T], with s the unknown scale
Matrix A = P.topLeftCorner(3, 3);
Vector a = P.col(3);
// do RQ decomposition to get s*K and cRw angles
Matrix sK;
Vector xyz;
boost::tie(sK, xyz) = RQ(A);
// Recover scale factor s and K
double s = sK(2, 2);
Matrix K = sK / s;
// Recover cRw itself, and its inverse
Rot3 cRw = Rot3::RzRyRx(xyz);
Rot3 wRc = cRw.inverse();
// Now, recover T from a = - s K cRw T = - A T
Vector T = -(A.inverse() * a);
return SimpleCamera(Pose3(wRc, T),
Cal3_S2(K(0, 0), K(1, 1), K(0, 1), K(0, 2), K(1, 2)));
}
namespace example1 {
const string filename = findExampleDataFile("5pointExample1.txt");
SfM_data data;
bool readOK = readBAL(filename, data);
Rot3 c1Rc2 = data.cameras[1].pose().rotation();
Point3 c1Tc2 = data.cameras[1].pose().translation();
PinholeCamera<Cal3_S2> camera2(data.cameras[1].pose(), Cal3_S2());
EssentialMatrix trueE(c1Rc2, Unit3(c1Tc2));
double baseline = 0.1; // actual baseline of the camera
Point2 pA(size_t i) {
return data.tracks[i].measurements[0].second;
}
Point2 pB(size_t i) {
return data.tracks[i].measurements[1].second;
}
Vector vA(size_t i) {
return EssentialMatrix::Homogeneous(pA(i));
}
Vector vB(size_t i) {
return EssentialMatrix::Homogeneous(pB(i));
}
//*************************************************************************
TEST (EssentialMatrixFactor, testData) {
CHECK(readOK);
// Check E matrix
Matrix expected(3, 3);
expected << 0, 0, 0, 0, 0, -0.1, 0.1, 0, 0;
Matrix aEb_matrix = skewSymmetric(c1Tc2.x(), c1Tc2.y(), c1Tc2.z())
* c1Rc2.matrix();
EXPECT(assert_equal(expected, aEb_matrix, 1e-8));
// Check some projections
EXPECT(assert_equal(Point2(0, 0), pA(0), 1e-8));
EXPECT(assert_equal(Point2(0, 0.1), pB(0), 1e-8));
EXPECT(assert_equal(Point2(0, -1), pA(4), 1e-8));
EXPECT(assert_equal(Point2(-1, 0.2), pB(4), 1e-8));
// Check homogeneous version
EXPECT(assert_equal(Vector3(-1, 0.2, 1), vB(4), 1e-8));
// Check epipolar constraint
for (size_t i = 0; i < 5; i++)
EXPECT_DOUBLES_EQUAL(0, vA(i).transpose() * aEb_matrix * vB(i), 1e-8);
// Check epipolar constraint
for (size_t i = 0; i < 5; i++)
EXPECT_DOUBLES_EQUAL(0, trueE.error(vA(i), vB(i)), 1e-7);
}
//*************************************************************************
TEST (EssentialMatrixFactor, factor) {
for (size_t i = 0; i < 5; i++) {
EssentialMatrixFactor factor(1, pA(i), pB(i), model1);
// Check evaluation
Vector expected(1);
expected << 0;
Matrix Hactual;
Vector actual = factor.evaluateError(trueE, Hactual);
EXPECT(assert_equal(expected, actual, 1e-7));
// Use numerical derivatives to calculate the expected Jacobian
Matrix Hexpected;
typedef Eigen::Matrix<double,1,1> Vector1;
Hexpected = numericalDerivative11<Vector1, EssentialMatrix>(
boost::bind(&EssentialMatrixFactor::evaluateError, &factor, _1,
boost::none), trueE);
// Verify the Jacobian is correct
EXPECT(assert_equal(Hexpected, Hactual, 1e-8));
}
}
//*************************************************************************
TEST (EssentialMatrixFactor, minimization) {
// Here we want to optimize directly on essential matrix constraints
// Yi Ma's algorithm (Ma01ijcv) is a bit cumbersome to implement,
// but GTSAM does the equivalent anyway, provided we give the right
// factors. In this case, the factors are the constraints.
// We start with a factor graph and add constraints to it
// Noise sigma is 1cm, assuming metric measurements
NonlinearFactorGraph graph;
for (size_t i = 0; i < 5; i++)
graph.add(EssentialMatrixFactor(1, pA(i), pB(i), model1));
// Check error at ground truth
Values truth;
truth.insert(1, trueE);
EXPECT_DOUBLES_EQUAL(0, graph.error(truth), 1e-8);
// Check error at initial estimate
Values initial;
EssentialMatrix initialE = trueE.retract(
(Vector(5) << 0.1, -0.1, 0.1, 0.1, -0.1).finished());
initial.insert(1, initialE);
#if defined(GTSAM_ROT3_EXPMAP) || defined(GTSAM_USE_QUATERNIONS)
EXPECT_DOUBLES_EQUAL(643.26, graph.error(initial), 1e-2);
#else
EXPECT_DOUBLES_EQUAL(639.84, graph.error(initial), 1e-2);
#endif
// Optimize
LevenbergMarquardtParams parameters;
LevenbergMarquardtOptimizer optimizer(graph, initial, parameters);
Values result = optimizer.optimize();
// Check result
EssentialMatrix actual = result.at<EssentialMatrix>(1);
EXPECT(assert_equal(trueE, actual, 1e-1));
// Check error at result
EXPECT_DOUBLES_EQUAL(0, graph.error(result), 1e-4);
// Check errors individually
for (size_t i = 0; i < 5; i++)
EXPECT_DOUBLES_EQUAL(0, actual.error(vA(i), vB(i)), 1e-6);
}
//*************************************************************************
TEST (EssentialMatrixFactor2, factor) {
for (size_t i = 0; i < 5; i++) {
EssentialMatrixFactor2 factor(100, i, pA(i), pB(i), model2);
// Check evaluation
Point3 P1 = data.tracks[i].p, P2 = data.cameras[1].pose().transform_to(P1);
const Point2 pi = PinholeBase::Project(P2);
Point2 reprojectionError(pi - pB(i));
Vector expected = reprojectionError.vector();
Matrix Hactual1, Hactual2;
double d(baseline / P1.z());
Vector actual = factor.evaluateError(trueE, d, Hactual1, Hactual2);
EXPECT(assert_equal(expected, actual, 1e-7));
// Use numerical derivatives to calculate the expected Jacobian
Matrix Hexpected1, Hexpected2;
boost::function<Vector(const EssentialMatrix&, double)> f = boost::bind(
&EssentialMatrixFactor2::evaluateError, &factor, _1, _2, boost::none,
boost::none);
Hexpected1 = numericalDerivative21<Vector2, EssentialMatrix, double>(f, trueE, d);
Hexpected2 = numericalDerivative22<Vector2, EssentialMatrix, double>(f, trueE, d);
// Verify the Jacobian is correct
EXPECT(assert_equal(Hexpected1, Hactual1, 1e-8));
EXPECT(assert_equal(Hexpected2, Hactual2, 1e-8));
}
}
//*************************************************************************
TEST (EssentialMatrixFactor2, minimization) {
// Here we want to optimize for E and inverse depths at the same time
// We start with a factor graph and add constraints to it
// Noise sigma is 1cm, assuming metric measurements
NonlinearFactorGraph graph;
for (size_t i = 0; i < 5; i++)
graph.add(EssentialMatrixFactor2(100, i, pA(i), pB(i), model2));
// Check error at ground truth
Values truth;
truth.insert(100, trueE);
for (size_t i = 0; i < 5; i++) {
Point3 P1 = data.tracks[i].p;
truth.insert(i, double(baseline / P1.z()));
}
EXPECT_DOUBLES_EQUAL(0, graph.error(truth), 1e-8);
// Optimize
LevenbergMarquardtParams parameters;
// parameters.setVerbosity("ERROR");
LevenbergMarquardtOptimizer optimizer(graph, truth, parameters);
Values result = optimizer.optimize();
// Check result
EssentialMatrix actual = result.at<EssentialMatrix>(100);
EXPECT(assert_equal(trueE, actual, 1e-1));
for (size_t i = 0; i < 5; i++)
EXPECT_DOUBLES_EQUAL(truth.at<double>(i), result.at<double>(i), 1e-1);
// Check error at result
EXPECT_DOUBLES_EQUAL(0, graph.error(result), 1e-4);
}
//*************************************************************************
// Below we want to optimize for an essential matrix specified in a
// body coordinate frame B which is rotated with respect to the camera
// frame C, via the rotation bRc.
// The "true E" in the body frame is then
EssentialMatrix bodyE = cRb.inverse() * trueE;
//*************************************************************************
TEST (EssentialMatrixFactor3, factor) {
for (size_t i = 0; i < 5; i++) {
EssentialMatrixFactor3 factor(100, i, pA(i), pB(i), cRb, model2);
// Check evaluation
Point3 P1 = data.tracks[i].p;
const Point2 pi = camera2.project(P1);
Point2 reprojectionError(pi - pB(i));
Vector expected = reprojectionError.vector();
Matrix Hactual1, Hactual2;
double d(baseline / P1.z());
Vector actual = factor.evaluateError(bodyE, d, Hactual1, Hactual2);
EXPECT(assert_equal(expected, actual, 1e-7));
// Use numerical derivatives to calculate the expected Jacobian
Matrix Hexpected1, Hexpected2;
boost::function<Vector(const EssentialMatrix&, double)> f = boost::bind(
&EssentialMatrixFactor3::evaluateError, &factor, _1, _2, boost::none,
boost::none);
Hexpected1 = numericalDerivative21<Vector2, EssentialMatrix, double>(f, bodyE, d);
Hexpected2 = numericalDerivative22<Vector2, EssentialMatrix, double>(f, bodyE, d);
// Verify the Jacobian is correct
EXPECT(assert_equal(Hexpected1, Hactual1, 1e-8));
EXPECT(assert_equal(Hexpected2, Hactual2, 1e-8));
}
}
//*************************************************************************
TEST (EssentialMatrixFactor3, minimization) {
// As before, we start with a factor graph and add constraints to it
NonlinearFactorGraph graph;
for (size_t i = 0; i < 5; i++)
// but now we specify the rotation bRc
graph.add(EssentialMatrixFactor3(100, i, pA(i), pB(i), cRb, model2));
// Check error at ground truth
Values truth;
truth.insert(100, bodyE);
for (size_t i = 0; i < 5; i++) {
Point3 P1 = data.tracks[i].p;
truth.insert(i, double(baseline / P1.z()));
}
EXPECT_DOUBLES_EQUAL(0, graph.error(truth), 1e-8);
// Optimize
LevenbergMarquardtParams parameters;
// parameters.setVerbosity("ERROR");
LevenbergMarquardtOptimizer optimizer(graph, truth, parameters);
Values result = optimizer.optimize();
// Check result
EssentialMatrix actual = result.at<EssentialMatrix>(100);
EXPECT(assert_equal(bodyE, actual, 1e-1));
for (size_t i = 0; i < 5; i++)
EXPECT_DOUBLES_EQUAL(truth.at<double>(i), result.at<double>(i), 1e-1);
// Check error at result
EXPECT_DOUBLES_EQUAL(0, graph.error(result), 1e-4);
}
} // namespace example1
